What people have said about working with us.
Drawn from clients across Singapore and APAC — covering proof-of-concept builds, evaluation engagements, and full model development projects.
← Back to HomeClient testimonials
Feedback from clients who engaged Nolux across different service types and sectors.
We brought Nolux in to evaluate a fraud detection model we'd built internally. The report they delivered was more thorough than we expected — they identified a specific edge case in how we'd partitioned the training data that was inflating our precision figures. That would have been a significant problem in production.
The proof of concept was exactly what we needed. We had a hypothesis about using ML for demand forecasting but weren't sure if our data was sufficient. Nolux ran the PoC, told us the data was marginal with our current volume, and gave us a clear recommendation to revisit once we had six more months of history. That's not the answer we wanted, but it was the right one.
The recommendation engine Nolux built for us is genuinely solid. What impressed me most was how clearly they scoped the work at the start — we had a very clear picture of what would and wouldn't be included before we committed. The handover documentation was detailed enough that our internal engineers picked it up without any confusion.
The team communicated very clearly throughout. When they ran into an unexpected issue with the label quality in our training set, they flagged it straight away rather than trying to work around it. That transparency saved us from building something that would have degraded quickly in production. The final model has been running in our environment for several months without issues.
I was expecting the evaluation to be a surface-level review, but it was far more detailed than that. They went through our feature engineering pipeline step by step and identified two places where data was leaking from the validation set. Fixing those two issues significantly changed the picture of how well our model was actually performing.
We started with a proof of concept to test whether our claims data could support a classification model. Nolux came back with a clear positive signal and a specific recommendation about what additional data would improve the model substantially. We followed that recommendation before moving to the full build, and it paid off. The whole process felt considered rather than rushed.
Three engagements in more detail
These case studies reflect real project patterns — identifying the challenge, the approach taken, and what the outcomes looked like.
A Singapore-based payments company had built a transaction risk scoring model in-house. Six months after deployment, the false positive rate was higher than expected. The team suspected the issue was in the training data but couldn't identify the cause.
Nolux reviewed the full training pipeline and tested the model against a held-out dataset. We identified that the temporal split used in training had inadvertently included future data in the training set, causing the model to learn patterns that weren't available at inference time.
After correcting the data split and retraining, the false positive rate dropped substantially. The team was also given a structured set of pipeline quality checks to prevent similar issues in future iterations. Delivered within the standard two-week timeline.
A Singapore-based online retailer wanted to improve product recommendations for returning customers. Their existing system was rule-based and not reflecting actual purchasing behaviour across their growing catalogue.
Following a data review workshop, we built a collaborative filtering model using the client's transaction and session history. The scope included data pipeline construction, model training, offline evaluation, and deployment guidance for their engineering team.
The recommendation model was deployed and integrated within three weeks of handover. The client's engineering team extended the feature set independently using the documentation provided. The approach reduced the overhead of the previous rule-based system significantly.
An insurance company wanted to explore whether historical claims data could predict the likelihood of a follow-up claim within twelve months. The business case was strong, but there was significant internal uncertainty about whether the data was sufficient.
We ran a PoC to test viability, which returned a positive signal with a specific recommendation: adding three additional data fields available internally would meaningfully improve predictive power. The client acted on that recommendation before commissioning the full build.
The PoC enabled a more informed build decision. The resulting model, trained with the additional data, performed substantially better than the initial prototype. Total investment across both phases was significantly lower than a single large-scope engagement would have been.
Speak to us about your project
A brief conversation is the easiest way to work out which engagement type suits your situation. No commitment involved.
Contact Us8 Shenton Way, #24-01
AXA Tower, Singapore 068811
Mon–Fri: 9:00 AM – 6:00 PM
Sat: 10:00 AM – 1:00 PM
Ready to add your project to the list?
Get in touch and we'll work out the right starting point together.
Get in Touch